---
language:
- en
license: llama3.1
library_name: transformers
pipeline_tag: text-generation
base_model: NousResearch/DeepHermes-3-Llama-3-8B-Preview
tags:
- astheral
- llama
- instruct
- qlora
---
# Astheral V3

Astheral is SEE7's 8.03B instruction model, built with Llama and fine-tuned from DeepHermes-3-Llama-3-8B-Preview.

[Chat](https://huggingface.co/spaces/SEE7/Astheral)

## Training

Merged BF16 weights from the V3 rank-16 QLoRA run: 41,943,040 adapter parameters, 1,713 training examples and 180 optimizer steps. Instruct and thinking modes are supported.

## Use

Load the merged checkpoint with Transformers in BF16, or use the hosted Space. The Space supports project instructions and up to 4,096 generated tokens.

## Reports and files

[Training and evaluation](https://huggingface.co/SEE7/Astheral/tree/main/v3-results) · [Extended experiments](https://huggingface.co/SEE7/Astheral/tree/v3-expanded-20261113)

Llama 3.1 licensing applies. See LICENSE, UPSTREAM_LLAMA3_LICENSE, NOTICE and the data attribution files.